Research Seminar on AI: Kinetic Theory for Metaheuristic Optimization Algorithms

 

Dienstag, 08.03.2022, 16.00 Uhr

The training, or optimization, of parameters plays a prominent role in the development of effective Artificial Intelligence models, but also in several other engineering tasks and scientific investigations. Despite their popularity among the community of practitioners, a class of optimization methods, the stochastic nature-inspired metaheuristic algorithms, has been usually neglected by the academic mathematical community due to their lack of interesting analytical properties (such as convergence guarantees).
The seminar will present a class of novel optimization algorithms, inspired by models studied in mathematical physics, which attempt to bridge this gap between the two communities. Indeed, while keeping the typical mechanisms of metaheuristic algorithms, these methods are amenable to theoretical understanding and convergence analysis. The Constrained Consensus-Based algorithm —studied by Giacomo Borghi and his team— will be presented as a case study on how to design and analyze metaheuristic algorithms under the framework on statistical mechanics

Giacomo Borghi is Ph.D. student of the Research Training Group “Energy Entropy and Dissipative Dynamics” at the RWTH Aachen University. He received his B.Sc. and M.Sc. in Mathematics both from the University of Bologna. In his doctoral project, he investigates kinetic models for particle-based optimization, together with his advisors Prof. Michael Herty and Prof. Lorenzo Pareschi.